Related Experiment Video
Updated: May 17, 2026
Reliability of Artificial Intelligence-Based Cone Beam Computed Tomography Integration with Digital Dental Images
Published on: February 23, 2024
A multi-task deep learning model for quality control of periapical radiographs: simultaneous technical error
Wen Fang1,2, Leqi Liu1,2, Xiaokan Wang1,2
1Stomatology Hospital, School of Stomatology, Zhejiang University School of Medicine, Zhejiang Provincial Clinical Research Center for Oral Diseases, Zhejiang Key Laboratory of Oral Biomedical Research, Hangzhou 310000, China.
Objectives:
To develop a multi-task deep learning model for the automated quality control of periapical radiographs, integrating identification of technical errors with standardized quality grading.
Methods:
A dataset of 3510 periapical radiographs was curated from 2 centers for model development and external validation. We developed a multi-task deep learning model to simultaneously perform: (1) multi-label classification of 6 common technical errors (positioning, processing, exposure, and angulation issues) and (2) hierarchical quality grading according to 2 established standards: the National Radiological Protection Board (NRPB: grades 1-3) and the College of General Dentistry (CGDent: ranks A-B). Model performance was evaluated using receiver operating characteristic-area under the curve (ROC-AUC), sensitivity, specificity, precision, accuracy, and F1-score on both internal and external test sets. Furthermore, the model's efficiency gain in daily quality control workflows was quantified.
Results:
Among 5 widely -adopted convolutional neural networks and transformer architectures evaluated, EfficientNet was selected as the optimal backbone. The final model identified technical errors with high performance on the external test set (sensitivity: 0.690-0.905; specificity: 0.965-0.995; AUC: 0.842-0.978). For quality grading, it achieved strong results for both trichotomous NRPB grading (sensitivity: 0.835-0.920; specificity: 0.881-0.989; AUC: 0.940-0.960) and binary ranking (sensitivity: 0.835-0.989; specificity: 0.835-0.989; AUC: 0.959).The system demonstrated superior computational efficiency, reducing the cumulative assessment time for a typical daily workload from over 32 minutes (manual review) to approximately 1.1 seconds. Grad-CAM visualizations confirmed the model's focus was aligned with technical errors, supporting clinical interpretability.
Conclusions:
This multi-task model efficiently automates radiographic quality control by identifying technical errors and assigning quality grades. It offers significant potential to standardize quality assurance and streamline dental workflows.